Combinatorial Learning of Robust Deep Graph Matching: An Embedding Based Approach.

Area of Science:

• Computer Vision
• Machine Learning
• Graph Theory

Background:

• Graph matching is crucial for establishing node correspondence but is computationally challenging (NP-hard).
• Accurate affinity modeling is essential for meaningful graph matching results, especially under noisy conditions.

Purpose of the Study:

• To develop an end-to-end deep neural network for learning graph matching.
• To address the NP-hard nature of graph matching by integrating deep learning techniques.
• To achieve robust and flexible graph matching adaptable to varying numbers of nodes and class-agnostic scenarios.

Main Methods:

• Utilized deep neural networks, including convolutional neural networks (CNNs) for feature extraction and graph neural networks (GNNs) for node embedding.
• Employed an end-to-end learning approach supervised by a combinatorial permutation loss.
• Developed an affinity kernel learning mechanism within the neural network architecture.

Main Results:

• Achieved state-of-the-art performance on extensive benchmark datasets.
• Demonstrated generalization capabilities across different categories and datasets.
• Showcased robustness against outliers and noise in graph matching tasks.

Conclusions:

• The proposed deep learning framework effectively addresses the challenges of graph matching.
• The end-to-end approach provides flexibility and robustness, outperforming existing methods.
• The method shows significant potential for real-world applications requiring accurate graph correspondence.